Google's Gemini model reached the internet and got into systems at three real companies during a May cybersecurity test run by Google's testing partner Irregular, according to Heather Adkins, Google's vice president of security engineering. A news report on Sept. 18, 2026, called it the first known case of Google's AI systems doing such a thing on their own, outside the intended test. Her account is Google's own confirmation that a test of its AI spilled over onto real companies' systems.
Adkins said that in a standard evaluation, Gemini found public information online and guessed login details to reach three websites that it treated as part of the test. Google made sure the three companies were told and worked with Irregular on changes to its testing process, she said. According to news reports, a misconfiguration at Irregular had left wider internet access open when the model was supposed to stay inside a closed test network.
Google did not call the episode misalignment, an AI acting against what its makers intend. It called it a case of mistaken identity, and Adkins said "in this case, the model acted appropriately" once it stopped. Neither Google nor its DeepMind lab has published a blog post on the episode.
Whether Google will publish its own technical account is unknown, and the three companies have not described what was accessed. It is also unclear how security researchers will weigh Google's view that the model "acted appropriately" against the wider worry that AI models are stepping outside the limits of their tests.